Rapid Classification and Quantification of Coal by Using Laser-Induced Breakdown Spectroscopy and Machine Learning

Coal is expected to be an important energy resource for some developing countries in the coming decades; thus, the rapid classification and qualification of coal quality has an important impact on the improvement in industrial production and the reduction in pollution emissions. The traditional methods for the proximate analysis of coal are time consuming and labor intensive, whose results will lag in the combustion condition of coal-fired boilers. However, laser-induced breakdown spectroscopy (LIBS) assisted with machine learning can meet the requirements of rapid detection and multi-element analysis of coal quality. In this work, 100 coal samples from 11 origins were divided into training, test, and prediction sets, and some clustering models, classification models, and regression models were established for the performance analysis in different application scenarios. Among them, clustering models can cluster coal samples into several clusterings only by coal spectra; classification models can classify coal with labels into different categories; and the regression model can give quantitative prediction results for proximate analysis indicators. Cross-validation was used to evaluate the model performance, which helped to select the optimal parameters for each model. The results showed that K-means clustering could effectively divide coal samples into four clusters that were similar within the class but different between classes; naive Bayesian classification can distinguish coal samples into different origins according to the probability distribution function, and its prediction accuracy could reach 0.967; and partial least squares regression can reduce the influence of multivariate collinearity on prediction, whose root mean square error of prediction for ash, volatile matter, and fixed carbon are 1.012%, 0.878%, and 1.409%, respectively. In this work, the built model provided a reference for the selection of machine learning methods for LIBS when applied to classification and qualification.